Learning to Optimize based on Unrolled ABIP
Yujie Peng, Ruoyu Diao, Liang Chen · Procedia Computer Science · 2025
Large-scale linear programming (LP) problems are ubiquitous in our daily lives. However, solving large-scale LP problems is far from trivial. Traditional LP algorithms perform well on small-scale problems but often encounter difficulties when solving large-scale LP problems. The recently developed ABIP algorithm has demonstrated outstanding performance. Although the penalty parameter update strategy in ABIP is crucial for reducing the number of iterations, the current penalty parameter updates in ABIP rely heavily on empirical heuristics. To address this, this paper employs an unrolling technique, introducing neural networks to dynamically adjust the penalty parameters in ABIP, thus proposing the LABIP algorithm. To obtain solutions with given precision, this paper adopts a two-stage framework: the framework begins with LABIP to generate an approximate solution, which is then re-fined to a relatively precise solution through ABIP. Experimental results demonstrate that using LABIP reduces the computational cost of ABIP by more than 20%.